Interpreting Log-Linked Count Coefficients

From a Log Coefficient to a Count-Mean Comparison

A coefficient in a count regression is not, by itself, a change in the expected count. In Poisson regression, coefficients enter the logarithm of the conditional mean. To interpret them, translate the log-scale comparison back to the mean-count scale. The translation is multiplicative: exponentiating a coefficient gives a ratio of conditional expected counts, and converting that ratio to a percentage gives the corresponding proportional change.

The word conditional is essential. A regression coefficient compares expected counts for specified covariate values, holding the other included covariates fixed. It does not automatically describe how an overall population average would change, because population averages also depend on which covariate values occur and how frequently.

Model and interpretation: Let $Y$ be a nonnegative count and let $X=(X_1,\ldots,X_k)$ denote its covariates. In Poisson regression, the conditional distribution is specified as $Y\mid X\sim\operatorname{Poisson}(\mu(X))$, with $$ \log\{\mu(X)\}=\log\{E(Y\mid X)\}=\beta_0+\beta_1X_1+\cdots+\beta_kX_k. $$ The conditional mean is therefore $\mu(X)=\exp(\beta_0+\sum_{j=1}^k\beta_jX_j)$. A coefficient describes a comparison of these conditional means, not a guaranteed change in an individual’s realized count.

The Poisson distribution assumption entails that the conditional variance equals the conditional mean. The coefficient translations below follow from the specified log-linear conditional mean. They do not require interpreting a fractional predicted mean as an observed count: a conditional expected count can be non-integer.

Incidence-Rate Ratios and Percentage Changes

For a one-unit increase in a continuous covariate $X_j$, holding all other covariates fixed, the linear predictor changes by $\beta_j$. Since the conditional mean is the exponential of that predictor, the new conditional mean is the old one multiplied by $\exp(\beta_j)$. This ratio is often called an incidence-rate ratio (IRR), including when the model concerns counts rather than rates per unit of exposure. No exposure offset is being considered here.

Definition: For a one-unit increase in $X_j$, the conditional mean ratio is $\exp(\beta_j)$. The exact percentage change in the conditional mean is $100\{\exp(\beta_j)-1\}\%$. For a change of $\Delta$ units, these become $\exp(\beta_j\Delta)$ and $100\{\exp(\beta_j\Delta)-1\}\%$, respectively, with the other covariates held fixed.

A positive coefficient gives an IRR above 1 and a positive percentage change; a negative coefficient gives an IRR below 1 and a negative percentage change. The percentage is generally not $100\beta_j$. That expression is an approximation when the coefficient is small, whereas exponentiation gives the exact percentage under the model. The unit of $X_j$ also matters: a coefficient per year, per dollar, or per one-point increase refers to different comparisons.

Worked Example: A Continuous Covariate and a Three-Unit Change

Suppose a hypothetical Poisson regression for annual clinic visits has $\hat\beta_{\text{age}}=0.12$, where age is measured in years. The other covariates are held fixed. Interpret a three-year increase, and calculate the predicted conditional mean if it is 2.40 visits per year at age 50.

A three-year increase changes the log conditional mean by $0.12\times3=0.36$. Its conditional mean ratio is $$ \exp(0.12\times3)=\exp(0.36)\approx1.433329. $$

Thus the model predicts a $100(1.433329-1)=43.33\%$ higher conditional mean for someone three years older, at the same values of the other included covariates. Starting from a conditional mean of 2.40 visits per year, the comparison mean is $$ 2.40\times1.433329\approx3.440 $$ visits per year. This is a model-based comparison of conditional means; it does not say that each person will have 1.040 additional visits.

Discrete Covariates: Compare the Values Directly

For a binary covariate coded 0 and 1, its coefficient compares the conditional mean when the covariate equals 1 with the conditional mean when it equals 0, at the same values of the other covariates. Exponentiating the coefficient gives the ratio for that comparison. This interpretation depends on the coding: if the reference category changes, the direction of the comparison changes too.

A ratio does not specify an absolute count difference by itself. The absolute difference depends on the reference conditional mean. The same IRR can correspond to a small difference at a low baseline mean and a larger difference at a high baseline mean.

Worked Example: A Binary Covariate and Its Reference Group

Consider a hypothetical model of annual workplace injury counts, with a binary indicator $D$ equal to 1 for a safety-training program and 0 for no program. Suppose $\hat\beta_D=-0.40$. At fixed values of the other covariates, the conditional mean for $D=1$ relative to $D=0$ is multiplied by $$ \exp(-0.40)\approx0.670320. $$

The exact percentage comparison is $100(0.670320-1)=-32.97\%$. In this model, the conditional expected injury count for the program group is 32.97% lower than for the reference group with the same included covariate values.

If the reference-group conditional mean at those covariate values is 4.80 injuries per year, the program-group conditional mean is $$ 4.80\times0.670320\approx3.218 $$ injuries per year. The conditional difference is $3.218-4.80=-1.582$ injuries per year. The ratio and percentage are fixed by the coefficient, but this absolute difference would be different at another reference mean. Because the model is observational as stated, this comparison alone does not establish that the program caused a reduction.

For a categorical covariate with more than two categories, coefficients are ordinarily interpreted relative to a specified reference category. For example, if a three-level factor is coded using indicator variables for its second and third levels, each coefficient compares that level with the omitted first level, holding other covariates fixed. To compare the second level directly with the third, subtract their coefficients on the log scale and exponentiate the difference. A coefficient must always be read together with the model’s coding and reference category.

Why Conditional Comparisons Are Not Population Changes

A conditional coefficient describes a comparison at fixed covariate values. An unconditional population mean instead averages over the covariate values in a population. Consequently, a conditional IRR cannot, by itself, tell us the difference between two groups’ observed overall mean counts: the groups may have different covariate compositions. Even if the model assigns the same conditional ratio at every covariate value, absolute conditional differences vary with the baseline mean.

To construct a model-implied population comparison, one must specify which covariate distribution is being averaged over. For example, one can predict the mean for each member of a defined population under each covariate value of interest and then average those predictions using the same population weights. That is a different quantity from the coefficient’s conditional comparison. It should be described as a standardized model-based comparison, and not automatically as a causal effect.

Worked Example: Averaging Conditional Predictions

Suppose a hypothetical model has a binary indicator $D$ and coefficient $\hat\beta_D=0.20$. Consider a specified population with two covariate profiles: 75% have a baseline conditional mean of 2.00 counts per period when $D=0$, and 25% have a baseline conditional mean of 8.00. These values represent differences in other covariates. Calculate the model-implied average at $D=0$ and at $D=1$, using the same profile weights for both predictions.

For either profile, changing $D$ from 0 to 1 multiplies the conditional mean by $$ \exp(0.20)\approx1.221403, $$ a 22.14% increase. At $D=0$, the weighted average prediction is $$ 0.75(2.00)+0.25(8.00)=3.500 $$ counts per period. At $D=1$, the two conditional predictions are $2.00(1.221403)=2.442806$ and $8.00(1.221403)=9.771222$. The corresponding weighted average is $$ 0.75(2.442806)+0.25(9.771222)\approx4.274910 $$ counts per period.

The model-implied average difference is therefore $4.274910-3.500=0.774910$ counts per period, or 22.14% of the baseline average. The numerical agreement in percentage occurs because this example applies the same IRR to every profile and uses the same weights in both predictions. The conditional absolute increases are not the same: they are about 0.443 and 1.771 counts per period. Changing the weights, or comparing groups with different covariate compositions, can produce a different overall comparison. These calculations are descriptive model predictions, not evidence of a causal effect.

A Short Derivation and Coefficient Uncertainty

The ratio interpretation follows directly from subtracting the log conditional means. Let $X_j$ change from $x$ to $x+\Delta$, while all other covariates remain fixed. The linear predictor at the new value differs by $\beta_j\Delta$. Exponentiating this difference gives the ratio of conditional means:

\begin{aligned} \log\{\mu(x+\Delta)\}-\log\{\mu(x)\} &=\beta_j(x+\Delta)-\beta_jx\\ &=\beta_j\Delta,\\ \frac{\mu(x+\Delta)}{\mu(x)} &=\exp\{\beta_j\Delta\}. \end{aligned}

Subtracting 1 from the ratio and multiplying by 100 converts it to a percentage change. This derivation also explains why the ratio is constant across values of the other covariates in the log-linear mean model, while the absolute difference $\mu(x)\{\exp(\beta_j\Delta)-1\}$ depends on the starting mean.

When a coefficient estimate is uncertain, its uncertainty should be carried through the transformation. For a one-unit comparison, a confidence interval for $\beta_j$ with endpoints $L$ and $U$ transforms to an interval for the IRR with endpoints $\exp(L)$ and $\exp(U)$. The corresponding percentage-change endpoints are $100\{\exp(L)-1\}\%$ and $100\{\exp(U)-1\}\%$. The endpoints are transformed in order because the exponential function is increasing. This transformation does not change the confidence level, provided the original interval has that coverage under its stated inferential conditions.

Common Mistakes

Reading the coefficient as a count difference. A coefficient of 0.12 does not mean 0.12 additional counts. Exponentiate it for a one-unit mean ratio, or exponentiate $\beta_j\Delta$ for a specified $\Delta$-unit comparison.
Calling $100\beta_j$ the exact percentage change. The exact change is $100\{\exp(\beta_j)-1\}\%$ for a one-unit increase. The linear approximation may be inaccurate for non-small coefficients.
Omitting the covariate conditions. State the comparison in context and say that other included covariates are held fixed. A regression coefficient is not generally an unconditional group comparison.
Equating a conditional association with a causal effect. Exponentiating a coefficient changes its scale; it does not supply a causal identification argument. Causal interpretation requires appropriate design and assumptions beyond the coefficient transformation.
Reporting only a percentage when an absolute difference is needed. The same percentage comparison can imply different numbers of counts at different baseline means. Give the relevant baseline and predicted means when an absolute comparison matters.

Key takeaway: In a log-linked count mean model, exponentiating a coefficient gives a conditional mean ratio. Convert that ratio to a percentage with $100(\text{ratio}-1)$, specify the covariate comparison and units, and do not mistake a conditional coefficient interpretation for an unconditional population change or a causal effect.

Check Your Understanding

For each question, distinguish the log-scale coefficient from the comparison on the conditional mean scale.

A fitted coefficient for a continuous covariate is 0.08 per unit. What is the exact percentage change in the conditional mean for a four-unit increase, holding other covariates fixed?
A binary indicator has coefficient $-0.25$, with 0 as the reference category. What is the conditional mean ratio for category 1 relative to category 0, and what is the percentage change?
Why does an IRR alone not determine the absolute difference in expected counts?
How can two groups’ unconditional average counts differ even when a model specifies a common conditional IRR across covariate profiles?
Why does a negative coefficient in an observational Poisson regression not, by itself, prove that changing the covariate would reduce counts?